LLM Regression Monitor

Overview

Automated behavioral regression monitoring for LLM apps. Captures baseline outputs, detects drift on a schedule, and fires WhatsApp or Slack alerts the moment something regresses.

Workflow Decision Tree

User request
├── "set up monitoring" / first time    → Full Setup (steps 1–5)
├── "run the monitor now"               → Step 4 only
├── "I changed my prompt/model"         → Step 3b (update baseline)
└── "configure alerts"                  → Step 5

Step 1 — Install

pip install llm-behave[semantic] pyyaml requests python-dotenv

Step 2 — Create test_suite.yaml

Create in the project root. Minimal example:

tests:
  - name: support_response
    prompt: "A customer says they never received their order. How do you respond?"
    provider: openai        # openai | anthropic | ollama | custom
    model: gpt-4o-mini
    assertions:
      - type: tone
        expected: "empathetic"
    drift:
      enabled: true
      threshold: 0.80

Set the API key for the chosen provider:

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...   # if using anthropic
# ollama needs no key

Read references/test-suite-format.md for the full field spec. Read references/providers.md for env vars and Ollama setup.

Step 3 — Capture Baselines

python scripts/capture_baseline.py

Saves ground-truth outputs to .llm_behave_baselines/. Run once before monitoring begins.

For more reliable baselines on important tests, set baseline_samples: 3 in test_suite.yaml — the script will call the LLM 3 times and save the most representative response, eliminating outliers.

To test real production behavior, add a system_prompt field to your test — it gets sent to the LLM exactly as in your app.

3b — Update after intentional prompt/model change

# Reset one test
python scripts/capture_baseline.py --update-baseline <test-name>

# Reset all
python scripts/capture_baseline.py --force

Step 4 — Run the Monitor

python scripts/run_monitor.py

Writes monitor_report.json. Exits 0 on all-pass, 1 on any failure (CI-compatible).

Step 4b — Track Trends (Predictive Alerts)

Run after each monitor run to log scores and detect gradual drift before it crosses the threshold:

python scripts/trend_tracker.py

Appends each run's scores to monitor_trend.jsonl. If a test's drift score has been declining for the last 5 runs and is within 0.10 of the threshold, it prints a warning. Exit code 2 = trend warning, 0 = all clear.

The daily schedule command (Step 6) already includes it:

python run_monitor.py; python trend_tracker.py; python send_alert.py

Step 5 — Configure Alerts

Set whichever channels you want in .env — scripts load it automatically.

# WhatsApp (requires wacli installed and logged in)
ALERT_WHATSAPP_TO="+1234567890"

# Slack
ALERT_SLACK_WEBHOOK="https://hooks.slack.com/services/..."

# Discord
ALERT_DISCORD_WEBHOOK="https://discord.com/api/webhooks/..."

# Email (Gmail example — use an App Password, not your main password)
ALERT_EMAIL_TO="you@example.com"
ALERT_EMAIL_FROM="sender@gmail.com"
ALERT_EMAIL_PASSWORD="your-app-password"
# Optional — defaults to smtp.gmail.com:587
# ALERT_EMAIL_SMTP="smtp.gmail.com"
# ALERT_EMAIL_PORT="587"

You can set any combination — all configured channels fire on failure. Send via:

python scripts/send_alert.py

Silent on green runs. Logs every alert to monitor_alerts.log regardless.

Step 6 — Schedule with OpenClaw Cron

Confirm the schedule with the user (default: 9am daily), then add:

• Schedule: 0 9 * * *
• Command: python run_monitor.py; python trend_tracker.py; python send_alert.py
• Directory: project root (where test_suite.yaml lives)

All three scripts run every day regardless of pass/fail. send_alert.py reads both monitor_report.json (hard failures) and trend_warnings.json (predictive warnings from trend_tracker.py) — it fires alerts only when there is something to report, and stays silent on fully green runs.

Common Errors